Deep End-to-End One-Class Classifier.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• One-class classification (OCC) is crucial for anomaly and outlier detection.
• General OCC is challenging due to data diversity and lack of novelty supervision.
• Existing methods struggle with end-to-end model design for detecting novelties.

Purpose of the Study:

• To propose an adversarial training approach for end-to-end trainable deep models for out-of-distribution detection.
• To develop a method that can learn the underlying distribution of the target class.
• To enhance the performance of novelty, anomaly, and outlier detection systems.

Main Methods:

• Jointly training two deep neural networks: a generator (R) and a discriminator (D).
• Utilizing the generator to create adversarial examples for characterizing the target class distribution.
• Employing the trained networks collaboratively for novelty detection during testing.

Main Results:

• Successfully learned the underlying distribution of the target class in experiments.
• Outperformed existing approaches on benchmark datasets for outlier detection (MNIST, Caltech-256).
• Demonstrated effectiveness in video anomaly detection tasks (UMN, UCSD datasets).

Conclusions:

• The proposed adversarial training method is effective for one-class classification.
• The approach enables end-to-end learning for detecting out-of-distribution samples.
• This method offers a robust solution for various anomaly and outlier detection applications.